Image parsing for image retrieval from large image data bases: from coloured image to coloured regions

This paper gives details of a series of low level image processing routines which successfully break an image into a set of coloured regions. The first stage in the process is multi-scale edge detection. A fixed set of different sized kernels are used with the results being put into a single ’edgeness’ image. A fixed lower threshold is applied to the edgeness image. A non-maximum suppression step is then applied. The histogram of edge strength in the non-max-suppressed image is used to set the high threshold for a hysteresis edge tracking routine. Points in the non-maxsuppressed image above the high threshold are used as seed points to grow edges with the full edgeness image being used as the search domain. The edge growing algorithm therefore suffers less from the topological damage resulting from nonmaximum suppression. A saliency filter is used to reject short crinkled edge chains. The second stage in the process uses the edge image to generate a series of Voronoi peaks. These are used as seed points for dilation type region growing. As the regions are grown the edgeness image is used to give a simple integrated measure of distance from a Voronoi centre. Pixels are then assigned membership in the ’nearest’ centre. Regions are not permitted to grow through edges in the edge image. A merging step is then applied to amalgamate regions with large non-edge shared boundary. The mean colour of each region is evaluated and the regions then form nodes in a graph.